New 'seed' input seeds sampling (torch.manual_seed before generation) so a fixed seed
reproduces the same output — no reroll when re-queueing while tweaking downstream nodes;
bump it to reroll (named 'seed' so the frontend adds control_after_generate). Generation
now honors ComfyUI cancel: a StoppingCriteria halts model.generate() promptly and
_check_interrupt() raises between passes. Workflows updated with the seed widget.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Labels crowded when splits are near each other. Now labels are shortened to S{n} and
skipped if within ~26px (canvas) / 30px (image) of the previous drawn label; notes gated
wider. Lines still drawn for every split.
Per clarification: the 721 group lines are clip markers, NOT segments. Segments come
ONLY from user split points, numbered globally from 1 — so the first split is seg1, not
seg23. 0 splits = 0 segments; N splits = N segments (seg k = point k-1 .. split k, tail
after the last split excluded). group lines still drawn (bold) + used for the group#.
JS matches (segPoints/nSegs). Empty summary now prompts to add splits.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Rebuilt around the real model: group_frames (721) are HARD splits; segments = user
splits UNION the group lines, so a segment never crosses a group boundary (last segment
in a group ends exactly on frame 721). segment_select is now a RANGE string ('A-B' /
'N' / '' = all) that crops waveform_image + audio + summary to segments A..B. Render
rewritten: mirrored waveform (uses top+bottom), bold group grid, thin segment lines,
labels along the top and notes along the bottom, selected range shaded. JS: guards
frame values (0 -> default, fixing the 400+-segments-on-reload explosion), writes only
USER splits to segments_json, dblclick=add split / shift-click=remove / click=seek.
group number fixed (boundary rounding). Workflows + README updated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Pivot to the user's model: segments are a fixed grid of subsegment_frames frames
(default 721 @ 24fps = one LTX clip), not arbitrary clicks. New inputs:
subsegment_frames (grid size) and segment_select (0=all, N=output ONLY chunk N —
crops waveform_image, AUDIO, and summary so you can generate/skip one beat at a time).
_render gains window-crop + frame markers + per-segment time/frame labels. JS rewritten:
draws the fixed grid, auto-fills the notes box with one segN: line per chunk (type or
dblclick to note), click-to-seek, playhead + time readout. Workflows updated for the
new widgets.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
An empty combo ([],) can't be built by ComfyUI, so the node wouldn't add to the
canvas. _audio_files() now always returns at least a placeholder entry.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The canvas Load expects the UI graph format (nodes[]+links[]+positions), not the
API format — so the _api.json came up empty. Added workflow_ltx_audio.json in UI
format (matched the schema from a real workflow: node dicts with pos/size/inputs/
outputs/widgets_values, link tuples, version 0.4), validated slots/links. Dropped
audio_upload from AudioWaveSegments so the widget layout is predictable (the JS
upload button stays). README points to the loadable file; _api.json kept for the bridge.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New workflow/workflow_ltx_audio_api.json: Audio Wave + Segments + LoadImage (subject)
+ Text (Prompt) holding the LTX audio system prompt -> Judge in chat mode
(json_output, model 3.5-9b) -> LTX beats JSON on the analysis output. Added a tiny
CalibratorText node ('Text (Prompt)', STRING source) needed to feed the judge's
forceInput system_prompt/user_prompt/axes sockets. README points to the workflow.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
SxCP is the Prompt-Builder pack's brand; these are Calibrator nodes. Renamed the
display names (cosmetic only — class_types unchanged, so no workflows/bridge break):
'External Prompt (Receptor)', 'Audio Prompt Guide', 'Audio Wave + Segments'; renamed
the JS extension id too. Kept the genuine Prompt-Builder (SxCP) references in the docs.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New node with a JS widget (web/audio_wave.js): upload an audio clip, play it, and click
the waveform to place segment boundaries (click add / drag move / dblclick note /
shift|right-click delete). Boundaries+notes serialize to a hidden segments_json that
drives Python segmentation (falls back to auto-split / notes syntax). Python node
(nodes/audio_wave_segments.py) loads the file (torchaudio/soundfile/librosa), builds
segments from the boundaries, and outputs waveform_image + audio_summary + AUDIO — same
contract as Audio Prompt Guide, so it feeds chat mode the same way. _attach_notes now
merges (keeps clicked notes). WEB_DIRECTORY re-enabled. JS is a first cut — needs testing
in ComfyUI (console logs on error); the Python node works standalone via segments_json/notes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The notes box now routes each line to a target: 'seg2: fast' (by number, also '2:'
/ 'S2:'), '10s-15s: drop' (by time range / '12.3s:'), or unprefixed = global. Segment
notes are attached to the matching segment in the summary (marked NOTE, overrides the
energy default) and drawn on the waveform image; time notes attach to overlapping
segments (strict end boundary); globals listed separately. README documents the syntax.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New node for audio-guided video prompts (LTX): takes a ComfyUI AUDIO clip + free-text
motion notes and outputs (1) a rendered energy-envelope IMAGE with beat/segment markers
so the vision model can see the audio shape, and (2) an audio_summary STRING with
duration, tempo/beats (librosa optional), per-segment energy + 8n+1-snapped frame counts
+ stage hints + the notes. Wire waveform_image -> Judge.reference_image and audio_summary
-> Judge.user_prompt (chat mode, json_output) for beat-aligned LTX JSON. librosa optional
(energy envelope + segments work without it). Registered in __init__; README workflow added.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
For JSON-producing system prompts (e.g. LTX prompt-relay), json_output=true pulls
the JSON object out of the reply (strips reasoning/prose/code-fences via _parse_json,
which handles nested schemas and reasoning-then-JSON) and returns it re-serialized;
falls back to raw text if none parses. agent_bridge gains --json-output.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
transformers .generate() is the slow path; reasoning token volume and swap_eval
(2 passes) are the multipliers. Now requests attn_implementation flash_attention_2
-> sdpa -> default automatically (free speedup, flash-attn optional). README gains
a Performance section: swap_eval off (biggest free win), flash-attn, smaller model/
fewer axes, avoid nf4 for speed, and vLLM/SGLang as the real production-speed path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Disabling thinking made reasoning models mark everything 'match' even when ref/gen
clearly differ. Added an enable_thinking toggle (default ON) threaded through the
generation path; the prompt now allows reasoning then asks for the result, and
verdict_rule explicitly warns against lazy 'match'. _parse_json now scans for the
JSON object AFTER the reasoning prose (last balanced object with 'axes'), and the
markdown fallback already reads reasoned per-axis output. Default max_new_tokens
2048->3072 so verdicts don't get cut off.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
axes, reference_description, system_prompt, user_prompt now render as INPUT
SOCKETS (forceInput) so they can be wired from other nodes — e.g. describe's
canonical output -> compare's reference_description, or a text node -> chat
prompts. Small config (report_dir, run_tag, model_path, ...) stays as typeable
fields. Unconnected sockets fall back to sensible defaults; the agent/bridge can
still set them by value via the API. Dropped the now-socket fields from the
example workflows; bumped their max_new_tokens to 2048.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Qwen3.5/3.6 are reasoning models — they 'think out loud' in markdown and never
reach the JSON, then get cut off at the token limit -> '(no parseable judgement)'.
Fixes: apply_chat_template(enable_thinking=False) + strip <think> blocks; hardened
'output ONLY JSON, do not think out loud' instruction; default max_new_tokens
1024->2048 (max 8192); and a markdown fallback parser (_parse_markdown_verdicts /
_parse_axes) that extracts per-axis {verdict,ref,gen} from the prose the model
reliably emits. describe falls back to using the raw text as the caption.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Widget-type inputs in the optional section render as connection sockets (not
editable boxes) in some ComfyUI frontends. Moved all widgets (report_dir, run_tag,
reference_description, system_prompt, user_prompt, keep_loaded, auto_download) into
required; only generated_image (a real node-to-node wire) stays optional. Same fix
for the receptor's source_file.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
fp16 offers nothing over bf16 for these models (removed from the quant dropdown;
loader still tolerant if passed). prompt_used was metadata-only — removed from the
node inputs, report payload/markdown, the bridge, and the example workflows.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New mode='chat' with system_prompt + user_prompt inputs runs your own prompt over
the image(s) and returns raw text in 'analysis' — reusing the same model dropdown,
quant, auto-download and backend. Makes it a one-node abliterated VLM for captioning,
tagging, Q&A, prompt-from-image, etc. agent_bridge gains --mode chat /
--system-prompt / --user-prompt (no receptor needed). Writes a chat report
(latest.json) for the agent. Docs updated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
No GGUF needed: huihui ships Qwen3.5-9B, Qwen3.6-27B, Qwen3.6-35B-A3B as multimodal
SAFETENSORS (abliterated), loadable via transformers AutoModelForMultimodalLM. Added
them to the model dropdown. _resolve_vl_classes now tries AutoModelForMultimodalLM
(3.5/3.6) and AutoModelForImageTextToText (Qwen3-VL) in name-based order, with
load fallback across candidates. model_select is now the model NAME only; precision
is the separate quant dropdown applied to it (repo_by_precision routes e.g. the local
fp8 dir). Aliases 3.5-9b/3.6-27b/3.6-35b. VRAM-by-quant table in README. Needs a
recent transformers for 3.5/3.6.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Per decision, this node stays transformers/safetensors. Removed the HauhauCS
GGUF presets from model_select; the manual .gguf guard now points to a dedicated
GGUF node (1038lab/ComfyUI-QwenVL, KLL535 Simple-Qwen3-VL-gguf). Dropdown lists
the huihui 4B/8B/30B-A3B judges with VRAM hints. Docs note GGUF models run in a
separate node and feed their text into the loop.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New model_select dropdown with suggested VRAM in each label: huihui Qwen3-VL
4B(local)/8B/30B-A3B (transformers, auto-download) and HauhauCS Qwen3.5-9B /
Qwen3.6-35B-A3B Uncensored Aggressive (GGUF). model_path is now the manual
override (empty = use dropdown). agent_bridge gains --model-select/--model-path.
The HauhauCS models are GGUF-only (no safetensors) so the transformers backend
can't load them yet — selecting one returns a clear 'GGUF backend pending'
message until the backend is added.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A discrete verdict collapses magnitude and a generic axis can hide what you're
calibrating (a blowjob where the head is 20cm away still reads sexual_act=oral ->
MATCH). New 'profile' input selects an act-specialized axis set (general / oral /
penetration / handjob / solo) whose act-critical axes capture distance explicitly
(mouth_genital_distance: touching/<5cm/10-20cm/>20cm, oral_depth, insertion_depth,
stroke_position, ...). axes now overrides the profile when set. agent_bridge gains
--profile; workflows + docs updated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Even the 30B mis-identifies named sex positions (doggy/cowgirl) from images, so
position_name is removed. The pose cluster is now purely observable geometry:
body_orientation enriched with facing direction (who faces whom), plus
limb_arrangement / contact_points / pose. The agent composes any named label from
these reliable primitives. 23 default axes. Docs/examples updated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The 4B over-uses 'partial' (mislabels identical ref/gen and clear opposites) and
also mis-identifies fine-grained content (e.g. names a position 'doggy'/'cowgirl'
when it is neither). Deterministic fix: force verdict=match when normalized
ref==gen. Prompt hardened to not default to 'partial' (opposites=mismatch). Docs:
the 4B is only reliable for coarse attributes — use the 30B for fine-grained
recognition; prefer grounded geometry axes over named-position labels.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Describe mode now produces a single coherent, internally-consistent canonical
scene description (paragraph + per-axis spec, written to canonical_reference in
the report). Compare gains an optional reference_description input: when set, it
anchors on that fixed text and shows only the generated image (no swap) — so the
reference side never drifts or self-contradicts across iterations; only the
generated image is re-described each turn. agent_bridge gains --ref-desc /
--ref-desc-file (reads the describe report's canonical_reference). Docs + example
workflow updated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The 4B's 0-1 scores were unreliable (identical ref/gen scored ~0.6), so the
judge now returns verdict match/partial/mismatch per axis; overall_score and a
new mismatch_count are computed from verdicts on our side (reliable, monotonic).
Expanded the action/pose cluster into position_name, body_orientation,
limb_arrangement, penetration, contact_points, genital_visibility (+ breast_size)
so explicit poses carry detail. Each axis now ships a one-line definition in the
prompt so gender_mix/subject_count stop absorbing positional text. 24 axes total.
Example workflows use the node default (axes=''). Docs realigned; stop condition
is now mismatch_count==0.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New mode on QwenVLImageJudge: 'describe' looks at the reference alone and returns
a prompt-ready caption + per-axis target spec to seed the very first prompt (the
generator has nothing to reproduce yet). 'compare' is the existing ref-vs-gen
scoring. generated_image is now optional (required only for compare); shared
generation refactored into _generate_from_messages; third output renamed
diff_analysis -> analysis (mode-agnostic). agent_bridge gains --mode (describe
needs no receptor/prompt); added workflow_describe_api.json. Docs updated with the
first-pass bootstrap step. Fixed error-return arity to 5-tuple.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The local VLM now only observes and scores; correction is left to the stronger
external agent. Each axis reports the target value (ref), the current value (gen)
and the closeness (score) — the target/current/distance an agent needs to
calibrate. Expanded to ~20 granular axes (identity/body/wardrobe/action/affect/
camera/render) so the action cluster stays discriminative for explicit content.
swap_eval now inverts ref/gen of the swapped pass; diff summary sorts worst-first;
default max_new_tokens 1024. Docs aligned.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ComfyUI-converted checkpoints ship the template as chat_template.jinja
(not on the processor), so apply_chat_template raised 'this processor does
not have a chat template'. Backfill processor.chat_template from
chat_template.jinja/.json or the tokenizer at load time, and fall back to a
hand-built Qwen-VL ChatML prompt if none exists. Also keep *.jinja in the
auto-download patterns.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
